feat: detect and parse q27 (signalnine/q27 engine) LLM backend

[PierpaoloPernici]

What

Adds q27 (signalnine/q27 engine) as a first-class LLM backend for the LLM probe, instead of mislabeling it as vLLM. Once detected, the LLM panel gets the full live-telemetry treatment: real-time decode/prefill tok/s, inference-health tiles, and backend-aware labels/badges.

Why

q27 exposes an OpenAI-compatible /v1/models plus Prometheus /metrics, so the probe fell through to the default vLLM classification — the badge read "vLLM" and every vLLM-specific tile stayed empty, because the vllm:* series it looks for don't exist on q27. (The q27 /metrics endpoint itself landed upstream in signalnine/q27#40; this PR teaches sparkDash to speak it.)

How

Detection (LlmProbe.js)

• /v1/models with owned_by: "q27", or
• Prometheus /metrics exposing q27_decode_tokens_total, q27_decode_tokens_processed_total, or q27_requests_total (regex anchored so q27_requests_errors_total and friends don't false-positive)
• Once classified q27, a failing /metrics no longer resets the backend to vLLM (same guard as ds4/exl3)

Metrics (_applyQ27Metrics)

• Sums across api= label sets (chat / completions / messages / responses), exactly like the ds4/vLLM paths
• Real-time tok/s from q27_*_processed counter deltas (they move during generation); falls back to the completion-based per-api totals for older binaries (step function: 0 while generating, jump at completion)
• Prefill is exact (per-request accounting, not vLLM estimates) and follows the ds4 convention: the main tile counts computed tokens only, so cache hits don't inflate the "real work" rate; the cached/uncached split doubles as the prefix-cache hit rate
• KV usage %, slots (inflight / total), TTFT / E2E / ITL p95 from q27_*_seconds histograms (with the +Inf == _count invariant check — mismatch refuses the quantile), and MTP/speculative accept ratio
• Engine state reports Active whenever the server is reachable (weights always resident, no sleep state), like the SGLang path
• Preemptions read 0 — q27 FIFO-queues and never preempts; the server exposes a constant-0 counter so the tile shows 0 instead of —
• No wait gauge — the Requests tile renders N run without the wait part

Frontend (LlmPanel.tsx, OverviewPage.tsx, types.ts)

• q27 badge/label everywhere backend names are rendered
• The inference-health tile grids render for vllm || q27; all tiles are null-safe
• Metric tooltips reworded backend-neutral ("from the engine's request history")

Decode benchmark (LlmStreaming.js)

• readServerGenerationTokens reads the q27 live processed counter (completion-total fallback), so bench results include server-side generation tok/s for q27 too

Also in this branch

• SystemCollector fix (a75af15): _readHostNetFile's fallback re-dispatched through _readHostFile, which routes /proc/net/* back into itself → Maximum call stack size exceeded on collectors without a host proc-netns bind mount. Now reads /proc/net directly. Also adds swap to the pseudo-fstype set so lsblk's [SWAP] pseudo-mount stops failing statfs with ENOENT on every poll.

Docs

• README: q27 in the feature table (auto-detect list, cached/uncached prefill backends), "LLM inference health" row (tiles now render for q27 too), and the LLM-probe backend list (also adds the missing EXL3 bullet)
• CHANGELOG: entry under [Unreleased] → Added

Testing

• 10 new unit tests in LlmProbe.q27.test.js: detection via owned_by and via /metrics (including processed-only expositions), gauge/counter parsing, idle → 0 tok/s, live processed counters driving real-time rates, cached-jump not inflating prefill, +Inf != _count refusing the quantile, and the full probe path (no vllm mislabel, context from max_model_len)
• Full suite: 173 tests, 172 passing (the 1 failure, showcasePrompts.test.js, fails identically on main — stale source assertion, unrelated)
• npm run typecheck ✅ · npm run build ✅

Screenshots

Overview — q27 auto-detected, live decode tok/s on the Spark card

Spark detail — full q27 LLM panel (tiles populated from q27_* metrics)

Live tok/s split by kind, KV cache, Requests reading 0 run without a wait gauge (FIFO), Preempts at 0 (never preempts), TTFT/E2E/ITL p95, prefix cache, MTP accept, and context from /v1/models max_model_len: